Description:[001] The present invention relates generally to the field of mineral identification and classification systems. More particularly, the invention concerns a computer-implemented hybrid framework that integrates handcrafted feature extraction methods, such as color histograms, texture descriptors, and shape analysis, with deep learning-based representations generated through convolutional neural networks. The invention further employs an optimized ensemble classifier for accurate recognition of mineral grains. The disclosed system is applicable across diverse domains including geosciences, mining, construction, material science, and environmental monitoring, where rapid, reliable, and automated identification of mineral compositions is required.
BACKGROUND OF THE INVENTION
[002] Mineral identification and classification represent essential processes in the fields of geology, mining, construction, and materials science. Traditional methods such as optical microscopy, Raman spectroscopy, and X-ray diffraction (XRD) are widely used for mineral characterization. While these techniques are accurate, they often require significant laboratory infrastructure, trained personnel, and extended processing time, which limit their use in high-throughput or field-based applications.
[003] Manual petrography, performed by expert mineralogists using thin-section analysis under optical microscopes, remains a common approach for identifying mineral grains. However, this process is inherently subjective and highly dependent on the skill and experience of the analyst. Human fatigue, inter-operator variability, and the complexity of mineral structures often lead to inconsistent and error-prone results, thereby reducing reliability in large-scale studies.
[004] Advanced methods, such as Scanning Electron Microscopy (SEM) combined with automated mineral analysis platforms (e.g., QEMSCAN, MLA, TIMA-X), have attempted to overcome the limitations of manual analysis. These systems can provide detailed grain-level information regarding mineral composition, shape, and abundance. However, their widespread adoption is restricted by high capital costs—often ranging from hundreds of thousands to millions of dollars—as well as the requirement for highly skilled operators and the long processing times associated with large sample volumes.
[005] Alternative approaches have emerged using image-based classification with computational techniques. Some studies rely on unsupervised clustering methods such as k-means or linear iterative clustering, while others have employed handcrafted features like texture descriptors (GLCM, LBP) for mineral grain recognition. These methods, while moderately effective, are constrained by limited generalization to real-world conditions, poor scalability, and a lack of integration with diverse feature modalities.
[006] The adoption of deep learning techniques in image recognition has transformed many domains, including medical imaging, agriculture, and autonomous systems. Convolutional Neural Networks (CNNs) have demonstrated high accuracy in recognizing complex visual patterns. In the context of mineral identification, CNNs can extract deep feature representations that capture subtle patterns not easily discernible by traditional feature engineering. However, reliance solely on deep features can result in reduced interpretability and vulnerability to class imbalance when datasets are limited or skewed.
[007] One of the major challenges in automated mineral identification is the availability of diverse, annotated datasets. Many existing datasets either cover a narrow range of mineral classes or fail to include the variability encountered in natural environments. This lack of comprehensive datasets has hindered the ability of machine learning models to generalize well in field deployments. For instance, similarities between minerals such as muscovite and biotite often result in high misclassification rates in standalone deep or handcrafted systems.
[008] In addition to dataset limitations, imbalanced class distributions present another challenge. Certain minerals are overrepresented in geological samples, while others appear rarely. This imbalance biases machine learning classifiers toward dominant classes, leading to poor recognition of minority minerals. Techniques such as Synthetic Minority Oversampling Technique (SMOTE) and WeightedRandomSampling have been proposed to address this issue, but integrating them effectively into a unified mineral identification pipeline remains a technical gap.
[009] There is also an unmet need for an interpretable yet highly accurate classification system. Handcrafted features such as shape descriptors and color histograms offer interpretability aligned with geological domain expertise, while deep features extracted via CNNs provide high-level abstractions. However, prior art largely treats these modalities separately, failing to exploit the potential synergy between traditional domain-driven features and data-driven deep learning representations.
[010] Thus, there exists a technical requirement for a hybrid mineral identification system that integrates handcrafted features with deep learning-based embeddings, addresses dataset imbalance, employs dimensionality reduction for computational efficiency, and leverages an optimized ensemble classifier. Such a system would overcome the drawbacks of traditional mineral identification techniques, reduce costs and dependency on expert operators, and enable real-time, reliable deployment across industrial and field applications.
SUMMARY OF THE INVENTION
[011] The present invention provides a computer-implemented hybrid mineral identification system designed to overcome the limitations of conventional analytical and machine learning techniques. Unlike traditional approaches that rely solely on either handcrafted descriptors or deep learning features, the invention integrates both modalities into a unified framework. By combining interpretable handcrafted descriptors with high-level deep representations, the system achieves superior accuracy, robustness, and interpretability in mineral grain classification.
[012] In one aspect, the invention employs a data preprocessing module that standardizes image dimensions, applies normalization, and performs augmentation operations such as cropping, flipping, affine transformations, and color jittering. This enhances the variability of training samples and significantly reduces overfitting risks, enabling the framework to generalize effectively across diverse mineral classes.
[013] In another aspect, the invention introduces a feature extraction unit capable of generating both handcrafted and deep features. Handcrafted features include HSV color histograms, Gray-Level Co-occurrence Matrix (GLCM) texture statistics, and Hu moment-based shape descriptors. In parallel, deep features are extracted from a fine-tuned ResNet50 convolutional neural network model, optimized for mineral datasets. The fusion of these features produces a rich and multi-modal representation of mineral grains.
[014] To address class imbalance inherent in mineral datasets, the invention integrates class balancing mechanisms. WeightedRandomSampling ensures balanced mini-batches during CNN training, while the Synthetic Minority Oversampling Technique (SMOTE) generates synthetic samples for underrepresented classes in traditional machine learning tasks. Together, these methods enable fair learning and reduce classifier bias toward majority classes.
[015] The fused feature vectors are further processed using dimensionality reduction techniques, including Principal Component Analysis (PCA), to retain 95% of the variance while discarding redundant and noisy dimensions. This results in faster computation, improved classifier efficiency, and reduced risk of overfitting.
[016] The classification stage of the invention employs an ensemble learning classifier based on XGBoost, which is optimized through hyperparameter tuning using Optuna or equivalent search frameworks. Parameters such as learning rate, tree depth, and subsampling ratio are systematically adjusted to achieve maximum predictive accuracy and generalization performance.
[017] The hybrid system achieves substantially improved performance over existing methods. Experimental validation demonstrates classification accuracies exceeding 97%, significantly outperforming standalone handcrafted models, deep models, and state-of-the-art approaches in the literature. The invention thus provides an unprecedented level of precision in distinguishing between mineral classes with similar morphological or textural features.
[018] In practical implementation, the invention may be deployed in field-portable devices, desktop systems, or cloud-based mineral analysis platforms, thereby enabling real-time identification of mineral grains during exploration surveys, mining operations, environmental assessments, and construction material evaluations. This versatility ensures broad industrial applicability and accessibility.
[019] Accordingly, the present invention delivers a novel hybrid mineral identification framework that combines domain knowledge with deep learning, incorporates class balancing and dimensionality reduction, and leverages optimized ensemble classifiers to provide reliable, scalable, and cost-effective mineral classification. The disclosed invention not only enhances operational efficiency but also reduces reliance on expert human judgment and high-cost laboratory instruments.
BRIEF DESCRIPTION OF THE DRAWINGS
[020] The accompanying figures included herein, and which form parts of the present invention, illustrate embodiments of the present invention, and work together with the present invention to illustrate the principles of the invention Figures:
[021] Figure 1, illustrates the block diagram of the hybrid mineral identification framework.
[022] Figure 2, depicts the feature extraction and fusion workflow in greater detail.
DETAILED DESCRIPTION OF THE INVENTION
[023] The present invention discloses a hybrid mineral identification framework that integrates handcrafted descriptors with deep learning-derived features to achieve high-accuracy mineral classification. The system comprises multiple modules, namely: a preprocessing unit, a handcrafted feature extraction unit, a deep learning feature extraction unit, a feature fusion module, an advanced feature processing module, and an optimized classifier. The modular design ensures flexibility, scalability, and suitability for both laboratory and field applications.
[024] The invention operates initially on input mineral images, obtained either from dedicated datasets such as MINET or directly from field-acquired imaging systems. The raw images are often variable in resolution, orientation, and illumination, which necessitates preprocessing to ensure uniformity. In this stage, each image is resized to 224 × 224 pixels to comply with the requirements of deep convolutional neural networks. Normalization is applied using ImageNet mean and standard deviation values to standardize pixel intensity ranges. Additionally, data augmentation techniques such as horizontal/vertical flipping, random cropping, affine transformations, color jittering, and perspective distortion are implemented to simulate real-world variability and enhance generalization capabilities of the classifier.
[025] The preprocessing module also incorporates class imbalance handling mechanisms. During training of the deep learning model, a WeightedRandomSampler is employed to oversample minority classes, ensuring that underrepresented minerals are proportionately included in each batch. For traditional machine learning tasks involving handcrafted features, the Synthetic Minority Oversampling Technique (SMOTE) generates synthetic examples for rare minerals, thereby balancing the dataset. This dual approach prevents bias toward dominant mineral classes and enhances the robustness of the classification framework.
[026] After preprocessing, the system performs handcrafted feature extraction to capture low-level, domain-relevant characteristics of mineral grains. Color descriptors are extracted in the HSV color space, which is more robust to illumination changes than RGB. Histograms across hue, saturation, and value channels capture distinctive mineral-specific patterns, such as the bright green of malachite or golden hues of pyrite. Texture descriptors are generated using Gray-Level Co-occurrence Matrix (GLCM) statistics, including contrast, dissimilarity, homogeneity, and energy, which describe the spatial relationships of pixel intensities. In addition, Hu moment-based shape descriptors are computed to capture geometric attributes such as elongation, circularity, and irregularity, which are useful for differentiating between flaky and granular minerals.
[027] In parallel, the deep learning feature extraction module processes mineral images using a fine-tuned ResNet50 convolutional neural network. The network, pretrained on ImageNet, is adapted to mineral classification tasks by replacing the final fully connected layer with a custom layer corresponding to the number of mineral classes. During training, the ResNet50 model learns to extract high-dimensional embeddings that capture subtle texture, structural, and spatial patterns beyond the reach of handcrafted features. The output of this process is a 2048-dimensional deep feature vector, representing each mineral image in an abstract feature space.
[028] The outputs of the handcrafted feature extractor and the deep learning feature extractor are combined in the feature fusion module. This module concatenates the handcrafted descriptors with the ResNet50 embeddings to generate a hybrid feature vector. The fusion ensures that low-level, interpretable cues (e.g., shape and color) are integrated with high-level abstract representations (e.g., learned CNN features). This synergy enhances the classifier’s ability to distinguish between minerals with similar textures or colors, such as muscovite and biotite, which often pose challenges in standalone classification approaches.
[029] To further refine the hybrid feature vectors, the invention incorporates an advanced feature processing stage. First, all features are standardized to have zero mean and unit variance, ensuring equal weighting during classification. Second, dimensionality reduction is applied using Principal Component Analysis (PCA), which compresses the feature space while retaining 95% of the variance. This step eliminates redundant dimensions, reduces computational burden, and minimizes the risk of overfitting, particularly in high-dimensional feature spaces.
[030] The processed hybrid features are then supplied to the classifier module, which employs an XGBoost ensemble learning algorithm. XGBoost was selected for its scalability, robustness, and ability to handle structured data with complex feature interactions. The classifier constructs boosted decision trees that learn non-linear relationships between fused features and mineral classes. To maximize predictive accuracy, hyperparameter optimization is performed using Optuna or equivalent frameworks. Parameters such as the number of estimators, tree depth, learning rate, subsampling ratios, and column sampling ratios are systematically tuned through iterative trials.
[031] The classification results are evaluated based on accuracy, precision, recall, and F1-score across all mineral classes. Experimental results indicate that the hybrid framework achieves classification accuracy of 97.44%, significantly surpassing models that use only handcrafted or only deep features. Furthermore, the hybrid system reduces misclassification between morphologically similar minerals, thereby ensuring robust performance in real-world applications.
[032] In terms of practical deployment, the invention is designed to be adaptable across multiple platforms. In one embodiment, the system may be embedded in field-portable mineral identification devices, enabling geologists and mining engineers to conduct rapid on-site analyses. In another embodiment, the system may be implemented in cloud-based petrography platforms, allowing remote mineral analysis from uploaded datasets. A further embodiment includes integration with desktop laboratory systems for high-throughput mineral classification in academic and industrial research facilities.
[033] The invention also allows future extensibility to incorporate multispectral and hyperspectral data in addition to RGB imaging. By integrating spectral data into the hybrid feature fusion pipeline, the classifier can achieve even higher accuracy and broader applicability. Furthermore, the inclusion of explainable AI techniques may enhance interpretability, providing domain experts with transparent insights into feature contributions during mineral identification.
[034] Overall, the invention discloses a novel hybrid mineral identification framework that combines handcrafted and deep learning features, addresses class imbalance, reduces computational inefficiency, and employs an optimized ensemble classifier for superior accuracy. The modular structure of the invention enables flexible deployment across diverse environments, making it a valuable advancement in mineral classification technology.
[035] The present invention provides a novel hybrid mineral identification framework that effectively integrates handcrafted descriptors with deep learning-derived features to achieve superior performance in automated mineral classification. By combining domain-specific features such as color, texture, and shape with abstract representations extracted by a fine-tuned ResNet50 model, the system delivers accuracy levels exceeding 97% on benchmark datasets. The incorporation of class balancing techniques, dimensionality reduction, and an optimized XGBoost classifier further enhances robustness, scalability, and efficiency, thereby overcoming the limitations of existing manual and computational approaches.
[036] One of the key advantages of this invention lies in its versatility of deployment. It can be applied in geological surveys, mining operations, construction material testing, and environmental monitoring, offering a low-cost, high-throughput, and reliable alternative to conventional laboratory-based methods. The modular design enables easy integration into portable field devices, desktop laboratory systems, or cloud-based analysis platforms, making it highly adaptable for diverse industrial and research contexts.
[037] The invention also opens avenues for significant future development and expansion. Integration with hyperspectral and multispectral imaging technologies will allow for more comprehensive spectral analysis of mineral grains, further improving classification accuracy. Similarly, embedding explainable artificial intelligence (XAI) techniques will enhance transparency and trust in model predictions, allowing domain experts to understand the contribution of different features in the classification decision-making process.
[038] Moreover, the system may be extended beyond mineral classification into related applications such as soil composition analysis, rock typing, sediment studies, and construction aggregate quality evaluation. By adapting the feature extraction modules to different imaging modalities, the framework can serve as a generalized platform for material characterization across multiple industries.
[039] In conclusion, the disclosed invention represents a substantial technical advancement in mineral identification by uniting interpretable handcrafted features with powerful deep learning models, supported by an optimized ensemble classifier. Its ability to operate at high accuracy, low cost, and with reduced dependence on expert human operators positions it as a transformative solution for geoscience and industrial applications. The inventive concept is not limited to the specific embodiments disclosed but may be practiced in various modifications and adaptations without departing from the spirit and scope of the invention.
 
, Claims:1. A computer-implemented mineral identification system, comprising:
o a preprocessing unit configured to normalize and augment mineral grain images;
o a handcrafted feature extraction unit configured to compute color histograms, texture descriptors, and shape descriptors;
o a deep learning module comprising a fine-tuned ResNet50 network for generating high-level feature embeddings;
o a feature fusion module configured to concatenate handcrafted features and deep embeddings into a hybrid feature vector; and
o a classification unit employing an optimized XGBoost ensemble classifier to predict mineral classes.
2. The system of claim 1, wherein the preprocessing unit applies data augmentation techniques including random cropping, flipping, affine transformations, perspective distortion, and color jittering to enhance dataset variability and reduce overfitting.
3. The system of claim 1, wherein the handcrafted feature extraction unit computes Gray-Level Co-occurrence Matrix (GLCM) descriptors, HSV color histograms, and Hu moment-based shape descriptors to represent domain-specific mineral characteristics.
4. The system of claim 1, wherein the deep learning module produces a 2048-dimensional embedding vector using a fine-tuned ResNet50 convolutional neural network, trained with a WeightedRandomSampler to address class imbalance.
5. The system of claim 1, wherein the feature fusion module generates a hybrid vector representation by combining handcrafted descriptors with deep embeddings, thereby improving recognition of visually similar minerals.
6. The system of claim 1, wherein the classification unit employs hyperparameter optimization using Optuna or equivalent frameworks to tune learning rate, tree depth, number of estimators, subsample ratio, and colsample_bytree for improved accuracy.
7. The system of claim 1, wherein the hybrid framework achieves a classification accuracy exceeding 95%, thereby surpassing conventional single-modality approaches.
8. The system of claim 1, wherein a dimensionality reduction unit applies Principal Component Analysis (PCA) to the hybrid feature vector to retain 95% of variance while reducing computational complexity.
9. A method for automated mineral identification, comprising the steps of:
o acquiring mineral grain images from a dataset or field device;
o preprocessing the images by resizing, normalization, and augmentation;
o extracting handcrafted features including color, texture, and shape descriptors;
o extracting deep features using a fine-tuned ResNet50 model;
o concatenating the features into a hybrid feature vector;
o applying dimensionality reduction and class balancing techniques; and
o classifying the mineral using an optimized XGBoost classifier.
10. The method of claim 9, wherein the system is deployed in portable field devices, desktop systems, or cloud-based platforms for real-time, high-throughput mineral classification in geological, mining, construction, and environmental applications.